Meta Superintelligence Labs runs on compute scarcity

Meta’s new “Superintelligence Labs” runs on a simple constraint: compute, not headcount. In an as-told-to essay for Business Insider, applied researcher Prakhar Agarwal says the work rhythm is set by long training and reinforcement-learning runs—often planned roughly 10 months out—followed by intense deadline periods as teams race to ship the next model iteration.

Agarwal’s description reads less like a conventional Big Tech org chart and more like a small, compute-gated workshop. With scarce GPU time, the usual corporate solution—hire more people and subdivide tasks—stops working, he argues, because additional staff simply dilutes the resource everyone needs to make progress. The result is a flatter structure, high-bandwidth communication, and a “fluid” definition of teams: researchers collaborate across lines depending on the problem, not the reporting chain.

That structure has second-order effects that matter outside Meta’s walls. If the primary bottleneck is compute, then internal status tends to follow who can justify scarce runs, not who can write the most documentation or shepherd the most process. Agarwal says much of the work is not documented and the code evolves faster than the docs, pushing researchers toward direct code archaeology and rapid iteration. In practice, that rewards people who can move quickly from “model fails at X” to a patch that improves the current model version before the next training cycle lands and potentially changes the problem.

It also shapes what “safety” looks like in day-to-day work. Agarwal says that when teams are further from a training deadline, they spend time on evaluations—trying to find failure modes in existing systems. But the underlying cadence remains deliverable-driven: if a fix misses the window, the next model version may render the effort irrelevant. That is a built-in bias toward interventions that can be validated quickly against today’s model, rather than slower work that only pays off when systems are deployed and exposed to messy real-world incentives.

Agarwal highlights a less visible advantage: frontier labs accumulate knowledge of what does not work. Papers record the successful recipe, he notes, but not the dozens of dead ends tried beforehand; internal teams retain that negative knowledge and build “intuition” from it. That is a competitive moat—but also a governance problem, because outsiders evaluating safety claims mostly see the published successes, not the internal failure catalog that determines what the lab will attempt next.

Meta is building an institution whose main scarce input is not talent but time on machines. The essay’s most concrete lesson is that, in this environment, the fastest way to influence outcomes is to control the next run.